inital softmax gpu kernel and tests added

[harz05]

A block of threads handles each row together, where a row is one slice along the reduction axis. Each thread strides over the row so the reads are coalesced. Row addressing is like- row_base + l*inner_stride. Exp and log go through alpaka::math.

The kernel uses online-softmax(https://arxiv.org/pdf/1805.02867). Each thread keeps a running (max, sum) pair over its slice in one fused pass, then the block combines the pairs with a single shared-memory tree reduction using the online merge operator, followed by a normalize pass.

Superseded:

First version: one thread per row, three passes, no shared memory.
2nd: block per row with two shared-memory tree reductions (max, then sum).

[harz05]

working on the warp approach to see if that could be done for it